ARIA: Adaptive Retrieval Intelligence Assistant -- A Multimodal RAG Framework for Domain-Specific Engineering Education

A research team from Johns Hopkins University has introduced ARIA (Adaptive Retrieval Intelligence Assistant), a novel multimodal Retrieval-Augmented Generation (RAG) framework designed to overcome the limitations of general-purpose LLMs in specialized education. To address issues like hallucinations and outdated knowledge, ARIA employs a sophisticated content extraction pipeline. It uses Docling for structured document analysis, Nougat for mathematical formula recognition, and the GPT-4 Vision API for interpreting diagrams, all powered by the e5-large-v2 embedding model for high semantic performance. This architecture allows it to accurately process the complex, multimodal materials typical of university-level engineering courses.

The system was rigorously evaluated using lecture content from a sophomore-level Statics and Mechanics of Materials course. Benchmarked against ChatGPT-5, ARIA demonstrated exceptional domain-specific performance. It achieved a 97.5% accuracy rate in filtering relevant questions, correctly answering all 20 relevant course queries while rejecting 58 out of 60 non-relevant ones. This resulted in a 90.9% precision score and a perfect 100% recall. Furthermore, its responses received an average quality rating of 4.89 out of 5.0. The researchers highlight that ARIA's course-agnostic design avoids the computational overhead of full model fine-tuning, presenting a scalable and efficient framework for deploying reliable, pedagogically consistent AI teaching assistants across various technical disciplines.